PEPA Based Performance Modeling for Robust Resource Allocations Amid Varying Processor Availability

Computational environments that are comprised of high performance parallel and/or distributed computing systems, often suffer performance degradation due to unforeseen variations in problem, algorithm, and system characteristics. In such environments, the underlying computational resource is considered to be non-dedicated. Therefore, to maintain a desired level of performance, there is a need for a study of the robustness of resource allocations. A robust resource allocation is expected to guarantee a predetermined level of performance, and optimize the execution performance, whenever possible. The goal of performance modeling and evaluation is to understand the behavior of the system (which includes the application and the computational system) and identify aspects of the system that are sensitive from a performance point of view. In general, much work has been focused on formulating robustness metrics for a number of resource allocation techniques via simulation, or using performance modeling for evaluating metrics, such as throughput and utilization in concurrent systems. Therefore, to the best of our knowledge, performance modeling for evaluating response times of resource allocations in parallel and distributed computing systems, and the related robustness analysis is still an open problem. The authors have validated the use of the performance evaluation process algebra (PEPA) for analyzing the robustness of static resource allocations for dedicated parallel computational resources in earlier research. In the work presented here, an Eclipse workbench that uses an implementation of PEPA, has been employed to numerically implement the model to obtain the performance of a number of static resource allocations obtained for a parallel execution environment with non-dedicated computational resources. The novelty of the proposed approach is to study the impact of varying processor/machine availability on the robustness of resource allocations. Moreover, as a novel approach, the variability in processor/machine availability is introduced into the performance modeling of the overall computational system. The performance is obtained via a numerical analysis of the performance of the modeled execution of applications on parallel computational resources with varying availability. Further, the results of the numerical analysis are used for evaluating the robustness of the static resource allocations. The work presented herein extends the benefits of using PEPA models towards a cost effective and low overhead analysis of robustness of resource allocations for a non-dedicated parallel computing system.